Automatic sleep scoring: a search for an optimal combination of measures

Objective

The objective of this study is to find the best set of characteristics of polysomnographic signals for the automatic classification of sleep stages.

Methods

A selection was made from 74 measures, including linear spectral measures, interdependency measures, and nonlinear measures of complexity that were computed for the all-night polysomnographic recordings of 20 healthy subjects. The adopted multidimensional analysis involved quadratic discriminant analysis, forward selection procedure, and selection by the best subset procedure. Two situations were considered: the use of four polysomnographic signals (EEG, EMG, EOG, and ECG) and the use of the EEG alone.

Results

For the given database, the best automatic sleep classifier achieved approximately an 81% agreement with the hypnograms of experts. The classifier was based on the next 14 features of polysomnographic signals: the ratio of powers in the beta and delta frequency range (EEG, channel C3), the fractal exponent (EMG), the variance (EOG), the absolute power in the sigma 1 band (EEG, C3), the relative power in the delta 2 band (EEG, O2), theta/gamma (EEG, C3), theta/alpha (EEG, O1), sigma/gamma (EEG, C4), the coherence in the delta 1 band (EEG, O1-O2), the entropy (EMG), the absolute theta 2 (EEG, Fp1), theta/alpha (EEG, Fp1), the sigma 2 coherence (EEG, O1-C3), and the zero-crossing rate (ECG); however, even with only four features, we could perform sleep scoring with a 74% accuracy, which is comparable to the inter-rater agreement between two independent specialists.

Conclusions

We have shown that 4-14 carefully selected polysomnographic features were sufficient for successful sleep scoring. The efficiency of the corresponding automatic classifiers was verified and conclusively demonstrated on all-night recordings from healthy adults.     

Circadian and homeostatic modulation of sleep in older adults during a 90-minute day study

STUDY OBJECTIVES: To identify age-associated changes in circadian and homeostatic characteristics of sleep in healthy elderly and young adults using electroencephalogram (EEG) power spectral analysis during a 90-minute sleep-wake schedule. DESIGN: Controlled clinical experiment. SETTING: University sleep laboratory. PARTICIPANTS: 16 older (77 +/- 5 years) and 19 younger adults (23 +/- 3 years). INTERVENTIONS: Subjects followed a 90-minute sleep-wake schedule (30 minutes in bed, 60 minutes awake) for 60 hours. Sleep was recorded for each bed-rest episode, and core body temperature was continuously recorded. The EEG power density was determined for non-rapid eye movement sleep in each bed-rest episode. Power density data were analyzed with mixed-effects models to assess rhythmic and linear components. RESULTS: Younger subjects had greater power in delta, theta, and sigma power bands across the study interval. Significant circadian rhythms were observed in delta, sigma, and beta power bands. Age-related differences in circadian modulation of EEG activity, indicated by significant interaction terms, were present in alpha and beta bands. A significant linear component was present in delta and theta power bands, with no significant age-interaction effect. CONCLUSIONS: Despite overall differences in the level of EEG power, older and younger adults exhibited similar rhythmic and linear patterns in most frequency bands. Age appears to affect circadian rhythmicity in higher EEG frequencies and homeostatic drive in lower EEG frequencies.     

Time‐varying correlations between delta EEG power and heart rate variability in midlife women: The SWAN Sleep Study

No studies have evaluated the dynamic, time‐varying relationship between delta electroencephalographic (EEG) sleep and high frequency heart rate variability (HF‐HRV) in women. Delta EEG and HF‐HRV were measured during sleep in 197 midlife women (M_age = 52.1, SD = 2.2). Delta EEG–HF‐HRV correlations in nonrapid eye movement (NREM) sleep were modeled as whole‐night averages and as continuous functions of time. The whole‐night delta EEG–HF‐HRV correlation was positive. The strongest correlations were observed during the first NREM sleep period preceding and following peak delta power. Time‐varying correlations between delta EEG–HF‐HRV were stronger in participants with sleep‐disordered breathing and self‐reported insomnia compared to healthy controls. The dynamic interplay between sleep and autonomic activity can be modeled across the night to examine within‐ and between‐participant differences including individuals with and without sleep disorders.     

Method for detection of respiratory cycle-related EEG changes in sleep-disordered breathing

STUDY OBJECTIVES: In sleep-disordered breathing (SDB), visual or computerized analysis of electroencephalogram (EEG) signals shows that disruption of sleep architecture occurs in association with apneas and hypopneas. We developed a new signal analysis algorithm to investigate whether brief changes in cortical activity can also occur with individual respiratory cycles. DESIGN: Retrospective. SETTING: University sleep laboratory. PARTICIPANTS: A 6 year-old boy with SDB. INTERVENTION: Polysomnography before and after clinically indicated adenotonsillectomy. MEASUREMENTS: For the first 3 hours of nocturnal sleep, a computer algorithm divided nonapneic respiratory cycles into 4 segments and, for each, computed mean EEG powers within delta, theta, alpha, sigma, and beta frequency ranges. Differences between segment-specific EEG powers were tested by analysis of variance. Respiratory cycle-related EEG changes (RCREC) were quantified. RESULTS: Preoperative RCREC were statistically significant in delta (P < .0001), theta (P < .001), and sigma (P < .0001) but not alpha or beta (P > .01) ranges. One year after the operation, RCREC in all ranges showed statistical significance (P < .01), but delta, theta, and sigma RCREC had decreased, whereas alpha and beta RCREC had increased. Preoperative RCREC also were demonstrated in a sequence of 101 breaths that contained no apneas or hypopneas (P < .0001). Several tested variations in the signal-analysis approach, including analysis of the entire nocturnal polysomnogram, did not meaningfully improve the significance of RCREC. CONCLUSIONS: In this child with SDB, the EEG varied with respiratory cycles to a quantifiable extent that changed after adenotonsillectomy. We speculate that RCREC may reflect brief but extremely numerous microarousals.     

A comparison of period amplitude analysis and FFT power spectral analysis of all-night human sleep EEG.

Zero-cross and zero-derivative period amplitude analysis (PAA) data were compared with power spectral analysis (PSA) data obtained with the fast Fourier transform in all-night sleep EEG from 10 subjects. Although PAA zero-cross-integrated amplitude showed good agreement with PSA power in 0.3-2 Hz, zero-cross analysis appears relatively ineffective in measuring 2-4 Hz and above waves. However, PAA zero-derivative measures of peak-trough amplitude correlated well with PSA power in 2-4 Hz. Thus, while PAA appears able to measure the entire EEG spectrum, the analytic technique should be changed from zero cross to zero derivative at about 2 Hz in human sleep EEG. PAA and PSA both demonstrate robust and interrelated across-night oscillations in three frequency bands: delta (0.3-4 Hz); sigma (12-16 Hz); and fast beta (20-10 Hz). The frequencies between delta and sigma, and between sigma and fast beta, did not show clear across-night oscillations using either method, and the two methods showed lower epoch-to-epoch agreement in these intermediate bands. The causes of this reduced agreement are not immediately clear, nor is it obvious which method gives more valid results. We believe that the three strongly oscillating frequency bands represent fundamental properties of the human sleep EEG that provide important clues to underlying physiological mechanisms. These mechanisms are more likely to be understood if their dynamic properties are preserved and measured naturalistically rather than being forced into arbitrary sleep stages or procrustean models. Both PAA and PSA can be employed for such naturalistic studies. PSA has the advantages of applying the same analytic method across the EEG spectrum and rests on more fully developed theory. Combined zero-cross and zero-derivative PAA demonstrates EEG oscillations that closely parallel those observed with spectral power, and the PAA measures do not rely on assumptions about the spectral composition of the signal. In addition, both PAA techniques can measure the relative contributions of wave amplitude and incidence to total power: These waveform characteristics represent different biological processes and respond differentially to a wide range of experimental conditions.     

Increased EEG sigma and beta power during NREM sleep in primary insomnia

The hyperarousal model of primary insomnia suggests that a deficit of attenuating arousal during sleep might cause the experience of non-restorative sleep. In the current study, we examined EEG spectral power values for standard frequency bands as indices of cortical arousal and sleep protecting mechanisms during sleep in 25 patients with primary insomnia and 29 good sleeper controls. Patients with primary insomnia demonstrated significantly elevated spectral power values in the EEG beta and sigma frequency band during NREM stage 2 sleep. No differences were observed in other frequency bands or during REM sleep. Based on prior studies suggesting that EEG beta activity represents a marker of cortical arousal and EEG sleep spindle (sigma) activity is an index of sleep protective mechanisms, our findings may provide further evidence for the concept that a simultaneous activation of wake-promoting and sleep-protecting neural activity patterns contributes to the experience of non-restorative sleep in primary insomnia.     

Increased EEG sigma and beta power during NREM sleep in primary insomnia

The hyperarousal model of primary insomnia suggests that a deficit of attenuating arousal during sleep might cause the experience of non-restorative sleep. In the current study, we examined EEG spectral power values for standard frequency bands as indices of cortical arousal and sleep protecting mechanisms during sleep in 25 patients with primary insomnia and 29 good sleeper controls. Patients with primary insomnia demonstrated significantly elevated spectral power values in the EEG beta and sigma frequency band during NREM stage 2 sleep. No differences were observed in other frequency bands or during REM sleep. Based on prior studies suggesting that EEG beta activity represents a marker of cortical arousal and EEG sleep spindle (sigma) activity is an index of sleep protective mechanisms, our findings may provide further evidence for the concept that a simultaneous activation of wake-promoting and sleep-protecting neural activity patterns contributes to the experience of non-restorative sleep in primary insomnia.     

DISCRIMINATION AMONG STATES OFCONSCIOUSNESS USING EEG SPECTRA

EEG recordings were made during waking (W) and the five sleep stages (REM, 1, 2, 3, and 4) on thirteen young adult males. For each stage, one-minute sections of the pa ietal EEG trace were digitized and subjected to Fourier analysis. The resulting spectral intensities were divided into five frequency bands; delta, theta, alpha, sigma, and beta. Linear discriminators for all six stages were calculated using stepwise multiple regression. The overall percent agreement with visual scoring was very poor, ranging from zero for stage 3 to 91% for stage 4. Linear discrimination between pairs of stages yielded slightly better results, but stages 1 and REM were indistinguishable. Delta is the best overall discriminator, increasing significantly through stages W, 1, 2, 3, and 4. Sigma is unique to sleep and is highest for stage 2. Theta is unimportant and beta plays no role at all. Spectral analysis of the parietal EEG lead is not sufficient to differentiate among the six states of consciousness studied here. The use of detectors for such phasic events as eye movement and K-complexes might aid sleep stage discrimination considerably.     

Influence of the menopausal transition on polysomnographic sleep characteristics: a longitudinal analysis

Study ObjectivesTo evaluate how change in menopausal status related to spectral analysis and polysomnographic measures of sleep characteristics.MethodsThe Study of Women’s Health Across the Nation (SWAN) Ancillary Sleep Study evaluated sleep characteristics of 159 women who were initially pre- or early perimenopausal and repeated the assessment about 3½ years later when 38 were pre- or early perimenopausal, 31 late perimenopausal, and 90 postmenopausal. Participants underwent in-home ambulatory polysomnography for two to three nights. Average EEG power in the delta and beta frequency bands was calculated during NREM and REM sleep, and sleep duration, wake after sleep onset (WASO), and apnea hypopnea index (AHI) were based on visually-scored sleep.ResultsThe women who transitioned to postmenopause had increased beta NREM EEG power at the second assessment, compared to women who remained pre-or early premenopausal; no other sleep measures varied by change in menopausal status. In multivariate models the associations remained; statistical controls for self-reported hot flashes did not explain findings. In secondary analysis, NREM beta power at the second assessment was greater among women who transitioned into the postmenopause after adjustments for initial NREM beta power.ConclusionsSleep duration and WASO did not vary by menopause transition group across assessments. Consistent with prior cross-sectional analysis, elevated beta EEG power in NREM sleep was apparent among women who transitioned to postmenopause, suggesting that independent of self-reported hot flashes, the menopausal transition is associated with physiological hyperarousal during sleep.     

Analysis of oscillatory patterns in the human sleep EEG using a novel detection algorithm

The different brain states during sleep are characterized by the occurrence of distinct oscillatory patterns such as spindles or delta waves. Using a new algorithm to detect oscillatory events in the electroencephalogram (EEG), we studied their properties and changes throughout the night. The present approach was based on the idea that the EEG may be described as a superposition of stochastically driven harmonic oscillators with damping and frequency varying in time. This idea was implemented by fitting autoregressive models to the EEG data. Oscillatory events were detected, whenever the damping of one or more frequencies was below a predefined threshold. Sleep EEG data of eight healthy young males were analyzed (four nights per subject). Oscillatory events occurred mainly in three frequency ranges, which correspond roughly to the classically defined delta (0-4.5 Hz), alpha (8-11.5 Hz) and sigma (11.5-16 Hz) bands. Their incidence showed small intra- but large inter-individual differences, in particular with respect to alpha events. The incidence and frequency of the events was characteristic for sleep stages and non-rapid eye movement (REM)-REM sleep cycles. The mean event frequency of delta and sigma (spindle) events decreased with the deepening of sleep. It was higher in the second half of the night compared with the first one for delta, alpha and sigma oscillations. The algorithm provides a general framework to detect and characterize oscillatory patterns in the EEG and similar signals.     

Evaluation of frequency and time-frequency spectral analysis of heart rate variability as a diagnostic marker of the sleep apnoea syndrome

The sleep apnoea/hypopnoea syndrome (SAHS) elicits a unique heart rate rhythm that may provide the basis for an effective screening tool. The study uses the receiver operator characteristic (ROC) to assess the diagnostic potential of spectral analysis of heart rate variability (HRV) using two methods, the discrete Fourier transform (DFT) and the discrete harmonic wavelet transform (DHWT). These two methods are compared over different sleep stages and spectral frequency bands. The HRV results are subsequently compared with those of the current screening method of oximetry. For both the DFT and the DHWT, the most diagnostically accurate frequency range for HRV spectral power calculations is found to be 0.019–0.036 Hz (denoted by AB₂). Using AB₂, 15 min sections of non-REM sleep data in 40 subjects produce ROC areas, for the DFT, DHWT and oximetry, of 0.94, 0.97 and 0.67, respectively. In REM sleep, ROC areas are 0.78, 0.79 and 0.71, respectively. In non-REM sleep, spectral analysis of HRV appears to be a significantly better indicator of the SAHS than the current screening method of oximetry, and, in REM sleep, it is comparable with oximetry. The advantage of the DHWT over the DFT is that it produces a greater time resolution and is computationally more efficient. The DHWT does not require the precondition of stationarity or interpolation of raw HRV data.     

Automated recognition of patients with obstructive sleep apnoea using wavelet-based features of electrocardiogram recordings

Patients with obstructive sleep apnoea syndrome (OSAS) are at increased risk of developing hypertension and other cardiovascular diseases. This paper explores the use of support vector machines (SVMs) for automated recognition of patients with OSAS types (±) using features extracted from nocturnal ECG recordings, and compares its performance with other classifiers. Features extracted from wavelet decomposition of heart rate variability (HRV) and ECG-derived respiration (EDR) signals of whole records (30 learning sets from physionet) are presented as inputs to train the SVM classifier to recognize OSAS± subjects. The optimal SVM parameter set is then determined by using a leave-one-out procedure. Independent test results have shown that an SVM using a subset of a selected combination of HRV and EDR features correctly recognized 30/30 of physionet test sets. In comparison, classification performance of K-nearest neighbour, probabilistic neural network, and linear discriminant classifiers on test data was lower. These results, therefore, demonstrate considerable potential in applying SVM in ECG-based screening and can aid sleep specialists in the initial assessment of patients with suspected OSAS.     

Regional differences of the sleep electroencephalogram in adolescents

The sleep electroencephalogram (EEG) was recorded from anterior (Fz/Cz) and posterior (Pz/Oz) bipolar derivations in two developmental groups: 20 pre- or early pubertal (Tanner 1/2, mean age 11.4 +/- 1.1 years, 11 boys) and 20 late pubertal or mature adolescents (Tanner 4/5, 14.1 +/- 1.3 years, 8 boys). A sleep-state independent reduction of EEG power over almost the entire frequency range was present in Tanner 4/5 compared with Tanner 1/2 adolescents. Spectral characteristics of the sleep EEG yielded state- and frequency-dependent regional differences that were similar in both developmental groups. Anterior predominance of power in delta and sigma ranges occurred in non-rapid eye movement sleep. Rapid eye movement sleep EEG power was greater in low delta, alpha, and sigma ranges for the posterior derivation and in theta and beta ranges for the anterior derivation. The decay rate of the sleep homeostatic process--reflected by the exponential decline of the 2-Hz EEG power band across the sleep episode--did not differ for derivations or groups. These results indicate that the nocturnal dynamics of sleep homeostasis are independent of derivation and remain stable across puberty.     

Pulse wave amplitude drops during sleep are reliable surrogate markers of changes in cortical activity

Background:During sleep, sudden drops in pulse wave amplitude (PWA) measured by pulse oximetry are commonly associated with simultaneous arousals and are thought to result from autonomic vasoconstriction. In the present study, we determine whether PWA drops were associated with changes in cortical activity as determined by EEG spectral analysis.Methods:A 20% decrease in PWA was chosen as a minimum for a drop. A total of 1085 PWA drops from 10 consecutive sleep recordings were analyzed. EEG spectral analysis was performed over 5 consecutive epochs of 5 seconds: 2 before, 1 during, and 2 after the PWA drop. EEG spectral analysis was performed over delta, theta, alpha, sigma, and beta frequency bands. Within each frequency band, power density was compared across the five 5-sec epochs. Presence or absence of visually scored EEG arousals were adjudicated by an investigator blinded to the PWA signal and considered associated with PWA drop if concomitant.Results:A significant increase in EEG power density in all EEG frequency bands was found during PWA drops (P < 0.001) compared to before and after drop. Even in the absence of visually scored arousals, PWA drops were associated with a significant increase in EEG power density (P < 0.001) in most frequency bands.Conclusions:Drops in PWA are associated with a significant increase in EEG power density, suggesting that these events can be used as a surrogate for changes in cortical activity during sleep. This approach may prove of value in scoring respiratory events on limited-channel (type III) portable monitors.Citation:Delessert A; Espa F; Rossetti A; Lavigne G; Tafti M; Heinzer R. Pulse wave amplitude drops during sleep are reliable surrogate markers of changes in cortical activity. SLEEP 2010;33(12):1687-1692.     

Diurnal variations in the waking EEG: comparisons with sleep latencies and subjective alertness

Daytime measures of sleep latency and subjective alertness do not correlate with one another, suggesting that they assess different aspects of alertness. In addition, their typical diurnal variations show very different time courses. Quantitative analysis of the waking electroencephalogram (EEG) has been proposed as an objective measure of alertness, but it is not clear how it compares with other measures. In this study, the waking EEG was measured in the daytime to determine the presence of diurnal variations in the activity of standard frequency bands and to compare these variations with the temporal patterns typical of sleep propensity and subjective alertness. Alertness was evaluated in four men and 12 women, aged 19-33 y. Assessments were conducted every 2 h, from 10.00 to 24.00, in the following order: a visual analogue scale of alertness, a waking EEG recording and a sleep latency test. The waking EEG was recorded with eyes open. For each recording session, 32-60 s of artefact-free signals were selected from the C3/A2 derivation, then subjected to amplitude spectral analysis. Four EEG frequency bands showed significant diurnal variations: delta, theta, sigma and beta1. None of these variations showed a significant correlation with the temporal patterns of sleep latencies or subjective alertness. At the individual level, however, theta band activity increased when subjective alertness decreased, suggesting that the theta band can be used to monitor variations in alertness in a given individual, even at the moderate levels of sleepiness experienced during the daytime.     

The EEG of the sleep onset period in insomnia: a discriminant analysis.

A period of rapid change in the wave components of the electroencephalogram (EEG) marks the transition from wake to sleep. Twenty-six insomniac and 28 control nights were studied in a discriminant analysis to determine whether this transitional state is modified in any way in subjects diagnosed for psychophysiological insomnia. A discriminant function was derived based on 20 insomniac and 22 normal nights. All 42 nights were correctly classified by this function. The sleep onset period, extending on the average over about 3 minutes, was characterized essentially by the beta and delta components of the EEG signal and by an activity index given by the ratio beta/delta, measured at the temporal lobe sites. Other variables included the subject's age and the magnitude of the changes occurring in the difference between activities in the right and left hemispheres. The variables contributing most to the discrimination were the activity index and beta, especially at the transitions from wake to stage 1 and from stage 1 to stage 2. The contribution of delta to the discrimination was less, but extended further in time to include stage 2 sleep. A test on the remaining six insomniac and six control nights gave a 75% classification accuracy, thus validating the derived discriminant function.     

Assessment of four statistical pattern recognition techniques to assist in obstructive sleep apnoea diagnosis from nocturnal oximetry

The aim of this study is to assess the utility of traditional statistical pattern recognition techniques to help in obstructive sleep apnoea (OSA) diagnosis. Classifiers based on quadratic (QDA) and linear (LDA) discriminant analysis, K-nearest neighbours (KNN) and logistic regression (LR) were evaluated. Spectral and nonlinear input features from oxygen saturation (SaO₂) signals were applied. A total of 187 recordings from patients suspected of suffering from OSA were available. This initial dataset was divided into training set (74 subjects) and test set (113 subjects). Twelve classification algorithms were developed by applying QDA, LDA, KNN and LR with spectral features, nonlinear features and combination of both groups. The performance of each algorithm was measured on the test set by means of classification accuracy and receiver operating characteristic (ROC) analysis. QDA, LDA and LR showed better classification capability than KNN. The classifier based on LDA with spectral features provided the best diagnostic ability with an accuracy of 87.61% (91.05% sensitivity and 82.61% specificity) and an area under the ROC curve (AROC) of 0.925. The proposed statistical pattern recognition techniques could be applied as an OSA screening tool.     

All night spectral analysis of EEG sleep in young adult and middle-aged male subjects

The sleep EEGs of 9 young adult males (age 20–28 years) and 8 middle-aged males (42–56 years) were analyzed by visual scoring and spectral analysis. In the middle-aged subjects power density in the delta, theta and sigma frequencies were attenuated as compared to the young subjects. In both age groups power density in the delta and theta frequencies declined from NREM period 1 to 3. In the sigma frequencies, however, no systematic changes in power density were observed over the sleep episode. In both age groups the decay of EEG power (0.75–7.0 Hz) over successive NREM-REM cycles and the time course of EEG power during NREM sleep was analyzed. The decay rate of both EEG power density over successive NREM-REM cycles and EEG power density during NREM sleep was smaller in the middle-aged subjects than in the young subjects. It is concluded that the age-related differences in human sleep EEG power spectra are not identical to the changes in EEG power spectra observed in the course of the sleep episode. Therefore age-related differences in EEG power spectra cannot be completely explained by assuming a reduced need for sleep in older subjects. The smaller decay rate of EEG power during NREM sleep in the middle-aged subjects is interpreted as a reduced sleep efficiency. The results are discussed in the frame work of the two-process model of sleep regulation.     

Dynamic changes in electroencephalogram spectral power with varying apnea duration in older adults

Sleep apnea elicits brain and physiological changes and its duration varies across the night. This study investigates the changes in the relative powers in electroencephalogram (EEG) frequency bands before and at apnea termination and as a function of apnea duration. The analysis was performed on 30 sleep records (375 apnea events) of older adults diagnosed with sleep apnea. Power spectral analysis centered on two 10‐s EEG epochs, before apnea termination (BAT) and after apnea termination (AAT), for each apnea event. The relative power changes in EEG frequency bands were compared with changes in apnea duration, defined as Short (between 10 and 20 s), Moderate (between 20 and 30 s) and Long (between 30 and 40 s). A significant reduction in EEG relative powers for lower frequency bands of alpha and sigma were observed for the Long compared to the Moderate and Short apnea duration groups at BAT, and reduction in relative theta, alpha and sigma powers for the Long compared to the Moderate and Short groups at AAT. The proportion of apnea events showed a significantly decreased trend with increased apnea duration for non‐rapid eye movement sleep but not rapid eye movement sleep. The proportion of central apnea events decreased with increased apnea duration, but not obstructive episodes. The findings suggest EEG arousal occurred both before and at apnea termination and these transient arousals were associated with a reduction in relative EEG powers of the low‐frequency bands: theta, alpha and sigma. The clinical implication is that these transient EEG arousals, without awakenings, are protective of sleep. Further studies with large datasets and different age groups are recommended.